Say the Posting's Words, Then Say What You Actually Did

Where Does the Keyword Idea Come From?

The idea has real sources, and one of them is a practice tool rather than an employer's scorer. Google's Interview Warmup let you speak an answer, transcribed it as you talked, and then reported which job-related terms and talking points the answer had covered ¹. That is a coverage report on your own rehearsal. Read as a description of hiring, it misleads.

Two things about that tool need keeping straight. The description above is Google's own copy, read from an archived snapshot of the page rather than from an audit. A coverage check on a rehearsal says nothing about how an employer's system works. The tool also appears to be gone. When checked on 1 September 2026, the page redirected to a general interview-tips article that never mentions Interview Warmup, so treat this as history rather than something to go and try ¹.

A check on your practice answer is not a description of an employer's process.

The confusion is easy to fall into, because coverage feedback is satisfying to receive. Cover the term, watch the term light up. Nothing in that loop tells you whether a person listening would believe you had done the thing the term names.

The habit it rewarded has outlived the tool.

What Does an AI Interview Scorer Actually Read?

One large vendor that publishes its method says the model reads the words and nothing else. HireVue's 2025 Explainability Statement describes a three-stage pipeline: speech-to-text transcription, then a natural-language model, then a score benchmarked against expert human rater evaluations of answers to similar competency-based questions ². What the score tracks, on that account, is how expert raters judged answers to the same kind of competency question.

Read that carefully, because it cuts both ways. It is a vendor describing its own product, which is not an audit, and it covers HireVue's AI-scored interviews only. The document calls itself a living document that will be updated, so it describes the 2025 configuration. Another vendor's system may read something else entirely. The statement also does not say the video goes unwatched. The recording exists and recruiters play it, and what a person notices there is a separate question from what the model reads ².

Benchmarked against human raters means the model's score is held up against what expert raters concluded.

Experienced interviewers reward an instance. On the employer's side the advice runs the same way, and running a structured interview about AI work means writing the rating levels as observable behaviour before the first interview. A level written that way asks what the candidate did and what came of it.

If four candidates in a row say cross-functional stakeholder management, the phrase separates none of them.

Why the Example Carries the Rating

The example carries the rating because the evidence favours the questions that ask for one. Past-behaviour means "tell me about a time when", and situational means "what would you do if". A meta-analysis comparing the two reported a substantially higher mean validity estimate for past-behaviour questions with descriptively anchored rating scales, .63 against .47 for situational ones scored the same way ³.

Faking is the second reason, and it was tested directly. An experiment inside the paper that built a scale for measuring interview faking found past-behaviour questions more resistant to faking than situational ones, and found that follow-up questioning increased faking ⁴. That was undergraduates, in person, in 2007. More resistant is relative rather than absolute, and the same paper reports that most candidates fake to some degree. Nothing in it tests a one-way video interview, which usually has no follow-up at all.

The validity numbers themselves are stale. They date from 2002, come from human-rated interviews with no AI scoring in them, and rest on range-restriction corrections that later work showed to be too generous, so the durable part is the comparison between formats rather than either value ³.

A format that is harder to bluff rewards the detail that is hardest to invent.

Employers hear the same problem from the other side. Every candidate now gives the same polished STAR answer because everyone pastes the same posting into the same assistant. The advice employers get is to ask for the option the candidate rejected, the person who disagreed, and the baseline under the number. Vocabulary answers none of those.

Rehearsed STAR answers fall flat for the same reason. The shape arrives and the specifics do not.

Use the Posting to Pick Which Story to Tell

The posting's most useful job in interview prep is casting rather than scripting. Read it for what it keeps returning to, then decide which of your real examples answers each likely question. If incident response appears in three separate bullets, prepare the incident. The shared vocabulary then arrives on its own, inside a sentence about something that happened.

On a resume, matching every keyword makes you median rather than shortlisted, and the interview version is stricter, because a person is listening for whether the term connects to anything you can describe.

There is some evidence that close matching counts for less once everyone can do it. On one online labour platform, after an AI cover-letter tool became available, the correlation between how closely a letter matched the job post and whether it drew a callback fell by 51% ⁵. Employers put weight on other signals instead, including prior work history. The study is a preprint using platform data and a difference-in-differences design, not an experiment and not corporate hiring. It is about letters rather than interviews.

The 51% is a fall in a correlation, not a fall in anyone's callback rate ⁵. What travels from it is the general shape: a signal everyone can produce tells people apart less well. Interview vocabulary is now that kind of signal, since any model will hand you the posting's phrases on request.

Borrowed words are cheap now. What you did is not.

Inventing the example instead has evidence against it. In a mock one-way video interview taken by 582 people across ten countries, honest tactics such as self-promotion were positively associated with rated performance ⁶. Deceptive ones, such as extensive image creation, were negatively associated with it. No job was at stake, the tactic use was self-reported, and humans did the rating, so that is an association rather than proof about any scoring model.

Take each phrase you meant to work in, and name the week you did it.

See how it works >